ON THE INDEPENDENCE NUMBER OF THE ERDOS-R˝ ENYI´ AND PROJECTIVE NORM GRAPHS AND A RELATED HYPERGRAPH
DHRUV MUBAYI∗ AND JASON WILLIFORD
Abstract. The Erd˝os-R´enyi and Projective Norm graphs are algebraically defined graphs that have proved useful in supplying constructions in extremal graph theory and Ramsey theory. Their eigenvalues have been computed and this yields an upper bound on their independence number. Here we show that in many cases, this upper bound is sharp in order of magnitude. Our result for the Erd˝os-R´enyi graph has the following reformulation: the maximum size of a family of mutually non-orthogonal lines in a vector space of dimension three over the finite field of order q is of order q3/2. We also prove that every subset of vertices of size greater than q2/2 + q3/2 + O(q) in the Erd˝os-R´enyi graph contains a triangle. This shows that an old construction of Parsons is asymptotically sharp. Several related results and open problems are provided.
1. Introduction The independence number α(G) of a graph or hypergraph G is the maximum size of a subset of vertices of G that contains no edge. The eigenvalues of the adjacency matrix of a graph provide an upper bound on its independence number. The aim of this paper is to examine the tightness of these bounds when applied to several well- known families of graphs. The particular graphs considered are the Erd˝os-R´enyi and Projective Norm graphs, and various subgraphs of these. We also consider a related hypergraph obtained from the Erd˝os-R´enyi graph that has recently proved useful in extremal hypergraph theory. Given a graph G and any set I ⊂ V (G) let e(I,I) be the number of ordered pairs of vertices in I which are adjacent. The aforementioned eigenvalue bound, found, for example, in [4], gives the following. Theorem 1. Let λ be the second largest eigenvalue in absolute value of the ad- jacency matrix of a d-regular graph G (possibly with loops) with n vertices. Then d 2 |e(I,I) − n |I| | ≤ λ|I|. In particular, if I contains no edges, except for possibly loops, then |I| ≤ (λ + 1)n/d.
1.1. The Projective Norm Graphs. Let t > 1 be a positive integer, Fq be the ∗ finite field of order q and Fq be the multiplicative group of Fq. The norm from 1+q+···+qt−2 Fqt−1 to Fq is the function N : Fqt−1 → Fq defined by N(X) = X .
∗Research supported in part by National Science Foundation grant DMS-0400812, and an Alfred P. Sloan Research Fellowship. 1 2 DHRUV MUBAYI∗ AND JASON WILLIFORD
Definition. Fix an integer t ≥ 2 and a prime power q. The Projective Norm graph ∗ G = Gq,t has vertex set V (G) = Fqt−1 × Fq with two vertices (A, a), (B, b) ∈ V (G) adjacent when N(A + B) = ab. For a graph G let ex(n, G) deonte the largest number of edges of any graph on n vertices which has no copy of G as a subgraph. Let Rk(G1,...,Gk) denote the smallest integer n such that any edge coloring in k colors of Kn must contain an i colored copy of Gi for some 1 ≤ i ≤ k, and let Rk(G) = Rk(G, . . . , G). The graphs Gq,t were first constructed by Alon, R´onyai and Szab´oto give improved lower bounds for the Tur´anfunction ex(n, Kt,s) and the Ramsey numbers Rk(Kt,s) where t ≥ 2 and s ≥ (t − 1)! + 1 (see [3]). In [22] T. Szab´ofound the eigenvalues of Gq,t to be ±q(t−1)/2, ±1, 0 and qt−1 −1. This was done independently by Alon and R¨odlin [2], who used these eigenvalues to provide tight bounds for Rk(Kt,s,...,Kt,s,Km) where again t ≥ 2 and s ≥ (t − 1)! + 1 (see [3]). t t−1 t−1 The projective norm graph Gq,t has n = q − q vertices, and degree q − 1. Using the results of [2, 22] for its eigenvalues and Theorem 1, we obtain for fixed t ≥ 2 (qt − qt−1)(q(t−1)/2 + 1) (1.1) α(G ) ≤ = (1 + o(1))q(t+1)/2 as q → ∞. q,t qt−1 − 1 Our first result shows that (1.1) gives the correct order of magnitude for all odd t. Theorem 2. Let t > 1 be an odd integer and q an odd prime power. Then q(t+1)/2 − q(t−1)/2 α(G ) ≥ . q,t 2 Thus as q → ∞, (t+1)/2 (t+1)/2 (1/2 + o(1))q < α(Gq,t) < (1 + o(1))q . The problem of determining whether (1.1) is sharp for even t seems to be more difficult. Open Problem 1. Find a construction of an independent set of size Cq(t+1)/2, C a constant, for even values of t > 2 or q a power of 2. 1.2. The Erd˝os-R´enyi Graph. We will pay special attention to the Projective Norm graphs in the case t = 2. The graph Gq,2 is a large induced subgraph of a well-known graph called the Erd˝os-R´enyi graph which we denote ERq. Since the graph ERq is of independent interest, we begin by explaining its construction and connection to Gq,2. Let q be a prime power and V be a 3-dimensional vector space over a finite field Fq. The projective geometry PG(2, q) is the triple (P, L, I) where P is the set of all 1-dimensional subspaces of V which we call points, L is the set of all 2-dimensional subspaces which we call lines, and the incidence relation I is containment. Points of PG(2, q) will be represented by left-normalized vectors, vectors whose left-most non-zero entry is 1, and which span the 1-dimensional space in question. Similarly lines will be represented by left-normalized vectors which span the orthogonal com- plement of the corresponding 2-dimensional subspace of V . We use round brackets for points and square brackets for lines to avoid confusion. Therefore a point (x0, x1, x2) is on a line [y0, y1, y2] if and only if x0y0 + x1y1 + x2y2 = 0. A polarity φ of a projective plane PG(2, q) is a bijective map from P ∪ L to itself that maps INDEPENDENCE NUMBER OF PROJECTIVE NORM AND RELATED GRAPHS 3 points to lines and lines to points and reverses incidence (meaning p ∈ l if and only if φ(l) ∈ φ(p)) and with the property that φ2 is the identity map on π. A point p is called absolute with respect to the polarity φ if p ∈ φ(p). The polarity graph of PG(2, q) with respect to a polarity φ is the simple graph (V,E) with vertex set equal to P such that for x, y ∈ P with x 6= y, {x, y} ∈ E if and only if x ∈ φ(y). (It should be noted that polarity graphs can alternately be defined with loops at each of the absolute points). The projective plane PG(2, q) is known to have the orthogonal polarity φo : (x , x , x ) 7→ [x , x , x ] and if q is a perfect square the unitary polarity φ : 0 1 2 √0 1√ 2 √ u q q q (x0, x1, x2) 7→ [x0 , x1 , x2 ]. Any other polarity of PG(2, q) is projectively equiv- alent to one of these forms (see [18]).
Definition. For q a prime power, the Erd˝os-R´enyi graph ERq is the orthogonal polarity graph of PG(2, q). Formally, the vertex set of ERq is the set of points of PG(2, q) with two distinct vertices (x0, x1, x2) and (y0, y1, y2) adjacent if and only o if x0y0 + x1y1 + x2y2 = 0. We define ERq to be the graph ERq with loops attached to the absolute points.
The graph ERq was introduced in this form by Erd˝osand R´enyi in [8] to give constructive examples of graphs with small maximum degree, relatively few edges and diameter 2. The graph ERq plays a notable role in extremal graph theory. Recall that if G is a graph, ex(n, G) denotes the greatest number of edges a graph on n vertices can have without containing G as a subgraph. Any graph on n vertices with ex(n, G) edges and which has no copy of G as a subgraph is called extremal. The asymptotic behavior of ex(n, G) is well understood if G is not bipartite; however, very little is known when G is bipartite (see [5] and [10] and the references therein). Of particular interest is the behavior of ex(n, C2k) where C2k denotes a cycle of length 2k. In [9] Erd˝os,R´enyi and S´osproved that ex(n, C4) is 1 3/2 asymptotic with 2 n using the graphs ERq for constructive lower bounds (this was done independently by Brown in [6]). F¨uredi later demonstrated in [12], [13] and [14] that the graphs ERq are extremal. Such exact results are relatively rare in extremal graph theory. The graph ERq has also been used to solve a similar problem for hypergraphs (see [19]). These hypergraphs will be dealt with in the last section. Many of our results require algebraic manipulations for which ERq is not suited. This leads us to the following ∗ Definition. For q an odd prime power, ERq is the graph whose vertex set is V (ERq) with two vertices (x0, x1, x2) and (y0, y1, y2) adjacent if and only if x0y2 − x1y1 + x2y0 = 0. Several of our calculations will be aided by the following fact:
∗ Proposition 3. The graph ERq is isomorphic to ERq.
The connection between the projective norm graph Gq,2 and ERq is given by the following Proposition.
Proposition 4. The graph Gq,2 is isomorphic to an induced subgraph of ERq. 2 2 ERq has q +q +1 vertices while Gq,2 has q −q. As Gq,2 is an induced subgraph of ERq missing 2q +1 of the vertices of ERq, their independence numbers are equal up to a linear (in q) error term. The question of determining the independence 4 DHRUV MUBAYI∗ AND JASON WILLIFORD number of ERq can be phrased as a simple question about vector spaces which seems independently interesting: What is the maximum number of mutually non-orthogonal 1-dimensional sub- spaces of a 3-dimensional vector space over Fq? o √ The eigenvalues of ERq are well known to be ± q, q + 1 (see [1]). It is clear o that the maximum size of a set of vertices of ERq with no edges except for possibly loops is equal to α(ER ). From Theorem 1, and noting that there are q + 1 loops q √ o 3/2 q+1 in ERq, we obtain α(ERq) ≤ q + q + q+1 . Since one of our results will be close to exact, we will introduce a bound due to Hoffman (see [17]) which is sharper than Theorem 1 with respect to lower order terms. However, this bound only bounds the size of the largest independent set which has no loops, so we must add the number of loops to the result. This bound states that if λ is the smallest eigenvalue of the adjacency matrix of a d-regular n-vertex graph G, then −nλ α(G) ≤ d − λ In the case of ERq we must add q + 1 to the bound, which after simplification gives us: 3/2 √ α(ERq) ≤ q + q + 1 Godsil and Newman generalize this bound if a restricted number of absolute points are allowed, see [15] for details. We will establish that the Hoffman bound gives the correct magnitude for α(ERq). Theorem 5. Let p be a prime, n a positive integer and q = pn. Then q3/2+q+2 2 for p odd, n even 3/2 120√q for p odd, n odd 73 73 α(ERq) ≥ 3/2 q√ for p = 2, n odd 2 2 √ q3/2 − q + q for p = 2, n even
3/2 In all cases, α(ER ) ≥ 120√q > .19239 q3/2. q 73 73
For even powers of 2 we can state exactly the size of the largest independent set which contains no absolute points. Theorem 6. Let q be an even power of 2. Then the size of the largest independent 3/2 √ set of ERq containing no absolute vertex is q − q + q. Given the level of precision in the previous result, it is natural to ask the follow- ing:
Open Problem 2. Construct an independent set I in ERq for all q which are not even powers of two such that |I| = q3/2 + O(q), or prove that no such set exists.
1.3. The Polarity Graph Uq. We next consider a graph closely related to ERq whose independence number can be found exactly for all q which are even powers of primes.
Definition. Let q be a square prime power. The unitary polarity graph Uq of PG(2, q) is the graph with vertex set V (ER ) with two vertices (x , x , x ) and √ q√ √ 0 1 2 q q q (y0, y1, y2) adjacent if and only if x0y0 + x1y1 + x2y2 = 0. INDEPENDENCE NUMBER OF PROJECTIVE NORM AND RELATED GRAPHS 5
As stated earlier, the graph Uq is the only other polarity graph which PG(2, q) admits. Uq is similar to ERq in that it has no 4-cycles and has diameter 2 (see [23] for a similar argument for ERq). However, it has fewer edges than ERq and so does not play the same role in extremal graph theory. Consequently, Uq is not as well known as ERq. As in ERq, there are absolute points. The graph formed − by deleting the absolute vertices of Uq will be denoted by Uq . Uq has precisely 3/2 q + 1 such absolute points, and these form an independent set in the graph Uq. We will show that this is the unique maximum independent set of Uq. This is in stark contrast to the graph ERq for which exact results remain elusive. To get the level of precision necessary in the upper bound of α(Uq) we will use a combinatorial bound as opposed to eigenvalue methods.
3/2 Theorem 7. Let q be a prime power. Then α(Uq) = q + 1, with the set of absolute points being the unique independent set of size q3/2 + 1.
It is interesting to note that the largest independent set in Uq consists of all the absolute points of Uq, while the best construction for ERq contains no absolute points. With the independence number of Uq resolved, it would be interesting to − see what is the magnitude of the largest independent set I of Uq (which contains no absolute points). Using the Hoffman bound and the fact that Uq has the same eigenvalues as ERq, we have the inequality
− 3/2 √ α(Uq ) ≤ q − q + q. From direct computation, this bound is not tight. For q = 4, 9 we have upper − − bounds of 6 and 21 while α(U4 ) = 4 and α(U9 ) = 19. A better upper bound and new constructive techniques are required to see if the size of I can also be found exactly as α(Uq).
1.4. The Erd˝os-R´enyi Hypergraph of Triangles. In [19] Lazebnik, and Ver- stra¨ete(using an idea of Lov´asz) construct a series of hypergraphs Hq of girth 5. These hypergraphs are used to determine the asymptotics of the Tur´annumber T3(n, 8, 4), defined as the maximum number of edges in a 3-graph on n vertices in which no set of 8 vertices spans more than 4 edges.
Definition. Let q be an odd prime power. Hq is the 3-graph whose vertex set is the set of nonabsolute points of V (ERq) and edge set is the set of triangles in ERq.
It is apparent from the construction of Hq that α(Hq) is the order of the largest triangle-free induced subgraph of ERq which contains no absolute points. In [21] q+1 Parsons constructs such a subgraph which has 2 vertices if q ≡ 3 (mod 4) and q 2 vertices if q ≡ 1 (mod 4) (Parsons’ aim was to study the automorphism groups of these subgraphs). We will show that Parsons’ construction is asymptotically tight using eigenvalue techniques.
2 3/2 Theorem 8. Let q be an odd prime power. Then α(Hq) ≤ q /2 + q + O(q). Thus in particular, 2 α(Hq) = (1/2 + o(1))q . Having shown that Parson’s graphs are asymptotically tight, we conjecture the following. 6 DHRUV MUBAYI∗ AND JASON WILLIFORD